Patty Coupeau is seeking a fellow for MSCA PF 2026 in graph neural networks explainability - me[...]

Organisation / Company Université d'Angers Department Direction de la Recherche - Cellule Europe Laboratory LARIS Is the Hosting related to staff position within a Research Infrastructure? No

The LARIS laboratory is a multidisciplinary research unit in Science and Technology belonging to the University of Angers. Our objective is to cultivate a strong disciplinary openness in the field of Information and Communication Sciences and Technologies (ICST). We encourage the development of scientific work that mainly relies on the skills of ICST, while associating, for certain themes or fields of application, those of Engineering Sciences and Life Sciences.

The unit is structured into three research teams focusing their research on the study of dynamic systems (SDO team), signal and image processing for life sciences (ISISV team) and reliabilty and operational performances of complex system (SFD team). This postdoctoral position is within the ISISV team.

The ISISV team includes researchers specializing in graph neural networks (Pr. Jean-Baptiste Fasquel, Dr. Nisrine Jrad, Dr. Patty Coupeau), in entropy on graphs and images (Pr. Anne Humeau-Heurtier) and researchers in the field of health offering several medical applications for our graph models. These include studies of the brain using MRI (Pr. Mickael Dinomais), EEG signals (Pr. Patrick Van Bogaert), walking analysis (Dr. Léna Carcreff) and motor skills evaluation (Dr. Josselin Demas, Dr. Pauline Ali).

Presentation of the project

Graph neural networks (GNNs) have attracted increasing interest due to their capacity to effectively model complex structured data, such as interactions between brain regions or parts of the human body. Each constituent element of the graph (nodes and edges with attributes) carries spatial, morphometric or connectivity properties that are considered important for understanding the problem of interest. Nevertheless, the methodological opacity of GNN models gives rise to fundamental questions regarding the understanding and reliability of their decisions.

In this context, the investigation of the explainability of graph-based models appears to be an essential area of research for fostering the development of GNN models in the medical world. Since 2019, numerous research projects have emerged, adapting explainability techniques specific to deep neural networks (DNNs) on images to graph neural networks (GNNs). These include gradient-based methods (Grad-CAM, Guided-BP), substitution methods (GraphLime, PGM-Explainer, GraphSVX), perturbation methods (GraphMask, GNNExplainer) and neural network-based methods (RCExplainer, CLEAR). The field still lacks comprehensive approaches that provide a unified analysis of the four essential components of graphs (nodes, edges, node attributes, and edge attributes). This represents a significant gap, particularly for attributed graphs, where edge characteristics convey important semantic information.

We are seeking candidates with experience in interpretability methods (e.g. applied to CNN, Transformers) who wish to learn about deep learning on graphs and contribute to medical research.

Keywords

Applicants must comply with the mobility rule: having stay in France less than 12 months in the past 3 years before the 9th of September 2026. Applicants also must have maximum 8 years of research experience after graduating their (first) PhD.

A dedicated support programme will be offered to the selected fellow with online training, webinars, proofreading as well as a potential funding of mobility in Angers during 3 days for dedicated writing sessions.